生成对抗网络及其应用研究综述

Review of Research on Generative Adversarial Networks and its Application

在人工智能引领的浪潮下,生成模型的理论研究与应用不断获得成功,已经成为当前研究的热点之一.因此,系统地研究生成模型及其应用具有重要的意义.本文首先将生成模型分为基于显式密度和基于隐式密度的生成模型,并介绍每类模型中具有代表的生成模型,分析它们的优缺点.其次,从应用层面分4类重点介绍了生成对抗网络在图像生成中的研究进展,即通过噪声生成图像、文本生成图像、图像到图像转换和交互式操控图像生成.然后从可解释性、可控性、稳定性和模型评价方法4个方面分析了生成对抗网络的理论研究进展.最后讨论了研究生成对抗网络潜在的突破口.

Under the tide led by artificial intelligence,the theoretical research and application of generative model have been successful and it has become one of the hot spots of current research.Therefore,it is of great significance to study the generative model and its application systematically.In this paper,the generative model is divided into the explicit density and the implicit density,then the representative generative models in the each type of model are introduced to analyze their advantages and disadvantages.Secondly,from the application level,this paper mainly introduces the research progress of generative adversarial networks in image generation,including image generation by noise,image generation by text,image to image translation and interactive controlling image generation.Then,the theoretical research progress of generative adversarial networks is analyzed from four aspects:interpretability,controllability,stability and model evaluation method.Finally,the potential breakthroughs of generative adversarial networks are discussed.